Seed Audio 1.0
BytePlus's all-round audio model — describe the sound, timbre, or delivery in natural language, or steer it with reference audio clips and a reference image. Audiobooks, voiceovers, game audio, ambient design; up to 2 minutes per run.
Capabilities
| Feature | Support |
|---|---|
| Text-to-Audio | Yes |
| Reference Audio | Up to 3 clips (max 30s each) — timbre & delivery-style guidance |
| Reference Image | 1 — the model reads the mood/scene |
| Voice Presets | Curated voice library (40 voices) |
| Max Duration | 120 seconds |
| Multi-purpose | Speech, SFX, ambience, music beds in one model |
Reference-Audio Guidance
Connect up to 3 audio clips (each ≤30s) and the model uses them as a style guide for the output — overall vocal character, pacing, and delivery:
Read in a delivery matching the references: "The expedition reached the
ridge at dawn, and what they saw below changed everything."
More reference material (up to the 3×30s budget) yields steadier style matching. For speech without reference audio, pick a voice from the built-in library instead.
Image-Guided Audio
Attach one image and Seed Audio scores it — connect a rainy street photo and ask for "the ambience of this scene". Useful for matching audio to a frame you've already generated.
One Model, Four Jobs
| Task | How to prompt |
|---|---|
| Voiceover / audiobook | Script in quotes + delivery notes ("warm, unhurried, slight smile") |
| Sound effects | Describe the event: "heavy oak door slams, stone hall reverb" |
| Ambience | Describe the space: "night market, distant chatter, sizzling wok" |
| Music bed | Genre + mood + instrumentation, like a music model |
Prompting Tips
- Direct the performance, not just the text. Pace, emotion, pauses: "reads slowly, breaking slightly on the last word".
- Keep reference clips clean. Isolated speech without music/noise gives much steadier style matching.
- Chunk long scripts. 120s per run — split chapters and keep the same references connected for a consistent voice across runs.
Limitations
- 120-second ceiling per generation
- Reference audio clips are capped at 30 seconds each
- No seed / determinism
See Also
- Gemini 3.1 Flash TTS / Qwen3 TTS — preset-voice narration
- ElevenLabs Sound Effects — dedicated SFX generation